metadata
language: ja
tags:
- audio
- automatic-speech-recognition
license: mit
library_name: ctranslate2
Whisper kotoba-whisper-v1.0 model for CTranslate2
This repository contains the conversion of kotoba-tech/kotoba-whisper-v1.0 to the CTranslate2 model format.
This model can be used in CTranslate2 or projects based on CTranslate2 such as faster-whisper.
Example
Install library and download sample audio.
pip install faster-whisper
wget https://huggingface.co/kotoba-tech/kotoba-whisper-v1.0-ggml/resolve/main/sample_ja_speech.wav
Inference with the kotoba-whisper-v1.0-faster.
from faster_whisper import WhisperModel
model = WhisperModel("kotoba-tech/kotoba-whisper-v1.0-faster")
segments, info = model.transcribe("sample_ja_speech.wav", language="ja", chunk_length=15, condition_on_previous_text=False)
for segment in segments:
print("[%.2fs -> %.2fs] %s" % (segment.start, segment.end, segment.text))
Benchmark
We measure the inference speed with four different Japanese speech audio on MacBook Pro with the following spec:
- Apple M2 Pro
- 32GB
- 14-inch, 2023
- OS Sonoma Version 14.4.1 (23E224)
| audio file | audio duration (min) | inference time (sec) |
|---|---|---|
| audio 1 | 50.3 | 2601 |
| audio 2 | 5.6 | 73 |
| audio 3 | 4.9 | 141 |
| audio 4 | 5.6 | 126 |
Conversion details
The original model was converted with the following command:
ct2-transformers-converter --model kotoba-tech/kotoba-whisper-v1.0 --output_dir kotoba-whisper-v1.0-faster \
--copy_files tokenizer.json preprocessor_config.json --quantization float16
Note that the model weights are saved in FP16. This type can be changed when the model is loaded using the compute_type option in CTranslate2.
More information
For more information about the original model, see its model card.